Missing values are decisions in disguise

A missing value describes the state of evidence, not the performance of an alternative. An absent cost estimate does not mean that implementation is free. An unresolved risk rating does not mean there is no risk. A blank readiness field does not establish incapacity. The comparison should preserve these distinctions. The canonical app rejects incomplete required numerical inputs and clears the current output until the problem is resolved. That behaviour is a substantive protection against presenting an apparently complete preference calculation on an incomplete evidence basis.

The initial response is diagnostic. Ask whether the value is missing because nobody gathered it, because the concept lacks a usable definition, because the estimate is being withheld or because the decision stage makes a precise value premature. Each cause suggests different work. A missing document may call for a source request. An unclear rubric requires methodological repair. A withheld estimate may require governance intervention. An early-stage option may need a deliberately bounded scenario rather than a false point estimate. A generic command to complete the spreadsheet does not identify these differences.

Missingness also has a perimeter. A sponsor may provide an estimate that omits integration, operating support or a necessary enabling asset. The numerical field is then populated but the decision input remains incomplete. Validation can confirm that a finite value is present; it cannot establish that the value covers the required scope. The evidence acceptance record should therefore check definitions and boundaries as well as numerical availability. An apparently complete dataset can be less informative than a visibly incomplete one if its exclusions are concealed.

The committee should predeclare how incomplete candidates will be handled. Options include deferring the candidate from the present comparison, requesting specified evidence, retaining it as a development option or running explicitly bounded scenarios under assumptions approved for the decision stage. These dispositions are different from assigning a convenient default. The record should explain what is missing, why the disposition is proportionate and what condition would bring the candidate back into the comparison. Otherwise omission can become an unreviewable way to remove an inconvenient proposal.

A default can be useful for demonstration, but it needs a clear identity. The app's reset values are labelled illustrative so a reader can explore the workflow. They are not estimates for the user's own proposal. A copied demonstration value should not silently become a project assumption. If an analyst elects to use an assumed value in a real early-stage scenario, the record should identify its origin, justification, sensitivity and owner. The decision process, rather than the existence of a populated field, determines whether the assumption is acceptable.

Treating missing performance as the least favourable score is a policy choice with consequences. It may encourage evidence completeness, but it can also penalise innovative or earlier-stage options relative to mature incumbents. That penalty may be appropriate for a near-term implementation decision and inappropriate for allocating investigation resources. The policy should be explicit about the commitment stage and the reason for the penalty. It should not be implemented secretly through a value that looks like measured poor performance.

Treating missing performance as average is another policy choice. It can preserve a candidate in a preliminary screen, but it grants it a position without evidence that the position is warranted. It may also alter the observed ranges and therefore the transformed scores of other candidates. The analyst should avoid presenting such a value as neutral. An assumed central position is still an assumption, with a potential effect on preference. The canonical app does not perform this imputation. Any separate scenario must be labelled and approved on its own terms.

Removing the criterion because one candidate lacks evidence can produce a different distortion. It changes the preference architecture for every candidate and may reward the very option that failed to supply the information. There may be legitimate reasons to simplify a preliminary screen when a criterion cannot yet be assessed consistently, but the committee should acknowledge the resulting blind spot. The original objective or constraint has not disappeared merely because the evidence is inconvenient. A later commitment may still require the missing assessment.

Removing the candidate is equally consequential. The observed-range transformation can change when a proposal leaves the set. The remaining proposals keep their raw performance but may receive different normalised values and scores. The analyst should therefore freeze and document the comparison universe after the missing-evidence disposition. A table should not mix scores computed with an incomplete candidate included and scores computed after it was removed. The reference set is part of the calculation's identity.

Bounded evidence scenarios can preserve an unresolved candidate without inventing certainty. The team states plausible interpretations of the disputed input, gives the basis for each bound and recomputes complete scenarios. This is useful when the bounds are credible and the resulting choice is decision-relevant. It is not permission to choose convenient extremes or call them a confidence interval. The bounds describe an authored uncertainty exercise unless supported by an appropriate estimation process. The evidence-conservative and evidence-optimistic fixtures later in the paper use exactly that limited interpretation.

A bound may remain too broad to support a commitment. If the preferred action changes across credible interpretations and the missing proposition can be investigated, the committee should consider delaying or staging the action. If the missing proposition cannot be resolved in time, the authorised decision maker may need to accept a documented uncertainty. The model should expose that choice. It should not manufacture a precise recommendation by collapsing an unresolved range into a single unexamined number.

The burden of evidence should be proportionate to consequence. A small reversible learning activity can tolerate a different level of estimation maturity from an irreversible implementation commitment. This does not mean that low-cost activities require no discipline. Their learning purpose still needs a clear question and disposition rule. It means that the evidence request should serve the present decision. Asking for implementation-grade detail before considering an exploratory option can eliminate learning opportunities without improving the immediate choice.

Diagnose a missing readiness rating by trying to complete the activity it is meant to describe. If the sponsor cannot name the operating task, the problem is an undefined scope; another readiness interview will not repair it. If the task is defined but the required capability has never been tested, a demonstration may resolve the proposition. If the capability is demonstrated but unavailable for the proposed period, the missing item is a resource commitment. These routes produce different work: define, test or secure. The evidence request should name that route so a document proving one proposition cannot silently close another.

An unresolved candidate can distort more than its own position. Under observed-range scaling, admitting an assumed extreme value may change the scores of supported candidates by moving an endpoint. Simply hiding the incomplete candidate in the final presentation therefore does not undo its influence. The diagnostic is to inspect whether its supplied placeholder defines any best or worst value, then distinguish two questions: whether the candidate belongs in the present comparison and whether the remaining scale still represents the intended tradeoffs. A development-only disposition should be applied before constructing the implementation comparison, with a fresh interpretation of any changed range.

The practical export check is a round trip between the numerical row and the missingness register. Select the candidate with an assumed readiness value, find the exact proposition that remains unsupported, and verify that the imported row still points to the same scoped activity. A valid input can reproduce an assumption perfectly while an accompanying note refers to a different design stage. Reject that join and repair the correspondence before reuse. Where the numerical schema cannot carry evidence maturity, the separate register supplies the link; it must distinguish an absent observation from an observed weak performance, since those imply different investigations.